# Stein's method

**Stein's method** is a general method in probability theory to obtain bounds on the distance between two probability distributions with respect to a probability metric. It was introduced by Charles Stein, who first published it in 1972,^{[1]} to obtain a bound between the distribution of a sum of -dependent sequence of random variables and a standard normal distribution in the Kolmogorov (uniform) metric and hence to prove not only a central limit theorem, but also bounds on the rates of convergence for the given metric.

## History

At the end of the 1960s, unsatisfied with the by-then known proofs of a specific central limit theorem, Charles Stein developed a new way of proving the theorem for his statistics lecture.^{[2]}
The seminal paper^{[1]} was presented in 1970 at the sixth Berkeley Symposium and published in the corresponding proceedings.

Later, his Ph.D. student Louis Chen Hsiao Yun modified the method so as to obtain approximation results for the Poisson distribution,^{[3]} therefore the method is often referred to as **Stein-Chen method**. Whereas moderate attention was given to the new method in the 70s, it has undergone major development in the 80s, where many important contributions were made and on which today's view of the method are largely based. Probably the most important contributions are the monograph by Stein (1986), where he presents his view of the method and the concept of *auxiliary randomisation*, in particular using *exchangeable pairs*, and the articles by Barbour (1988) and Götze (1991), who introduced the so-called *generator interpretation*, which made it possible to easily adapt the method to many other probability distributions. An important contribution was also an article by Bolthausen (1984) on a long-standing open problem around the so-called *combinatorial central limit theorem*, which surely helped the method to become widely known. {{ safesubst:#invoke:Unsubst||date=__DATE__ |$B=
{{#invoke:Category handler|main}}{{#invoke:Category handler|main}}^{[citation needed]}
}}

In the 1990s the method was adapted to a variety of distributions, such as Gaussian processes by Barbour (1990), the binomial distribution by Ehm (1991), Poisson processes by Barbour and Brown (1992), the Gamma distribution by Luk (1994), and many others.

## The basic approach

### Probability metrics

Stein's method is a way to bound the distance of two probability distributions in a specific probability metric. To be tractable with the method, the metric must be given in the form

Here, and are probability measures on a measurable space , and are random variables with distribution and respectively, is the usual expectation operator and is a set of functions from to the real numbers. This set has to be large enough, so that the above definition indeed yields a metric. Important examples are the total variation metric, where we let consist of all the indicator functions of measurable sets, the Kolmogorov (uniform) metric for probability measures on the real numbers, where we consider all the half-line indicator functions, and the Lipschitz (first order Wasserstein; Kantorovich) metric, where the underlying space is itself a metric space and we take the set to be all Lipschitz-continuous functions with Lipschitz-constant 1. However, note that not every metric can be represented in the form (1.1).

In what follows we think of as a complicated distribution (e.g. a sum of dependent random variables), which we want to approximate by a much simpler and tractable distribution (e.g. the standard normal distribution to obtain a central limit theorem).

### The Stein operator

We assume now that the distribution is a fixed distribution; in what follows we shall in particular consider the case where is the standard normal distribution, which serves as a classical example of the application of Stein's method.

First of all, we need an operator which acts on functions from to the real numbers, and which 'characterizes' the distribution in the sense that the following equivalence holds:

We call such an operator the *Stein operator*. For the standard normal distribution, Stein's lemma exactly yields such an operator:

thus we can take

We note that there are in general infinitely many such operators and it still remains an open question, which one to choose. However, it seems that for many distributions there is a particular *good* one, like (2.3) for the normal distribution.

There are different ways to find Stein operators. But by far the most important one is via generators. This approach was, as already mentioned, introduced by Barbour and Götze. Assume that is a (homogeneous) continuous time Markov process taking values in . If has the stationary distribution it is easy to see that, if is the generator of , we have for a large set of functions . Thus, generators are natural candidates for Stein operators and this approach will also help us for later computations.

### Setting up the Stein equation

Observe now that saying that is close to with respect to is equivalent to saying that the difference of expectations in (1.1) is close to 0, and indeed if it is equal to 0. We hope now that the operator exhibits the same behavior: clearly if we have and hopefully if we have .

To make this statement rigorous we could find a function , such that, for a given function ,

so that the behavior of the right hand side is reproduced by the operator and . However, this equation is too general. We solve instead the more specific equation

which is called *Stein equation*. Replacing by and taking expectation with respect to , we are back to (3.1), which is what we effectively want. Now all the effort is worth only if the left hand side of (3.1) is easier to bound than the right hand side. This is, surprisingly, often the case.

If is the standard normal distribution and we use (2.3), the corresponding Stein equation is

which is just an ordinary differential equation.

### Solving the Stein equation

Now, in general, we cannot say much about how the equation (3.2) is to be solved. However, there are important cases, where we can.

*Analytic methods*. We see from (3.3) that equation (3.2) can in particular be a differential equation (if is concentrated on the integers, it will often turn out to be a difference equation). As there are many methods available to treat such equations, we can use them to solve the equation. For example, (3.3) can be easily solved explicitly:

*Generator method*. If is the generator of a Markov process as explained before, we can give a general solution to (3.2):

where denotes expectation with respect to the process being started in . However, one still has to prove that the solution (4.2) exists for all desired functions .

### Properties of the solution to the Stein equation

After showing the existence of a solution to (3.2) we can now try to analyze its properties. Usually, one tries to give bounds on and its derivatives (which has to be carefully defined if is a more complicated space) or differences in terms of and its derivatives or differences, that is, inequalities of the form

for some specific (typically or , respectively, depending on the form of the Stein operator) and where often is taken to be the supremum norm. Here, denotes the differential operator, but in discrete settings it usually refers to a difference operator. The constants may contain the parameters of the distribution . If there are any, they are often referred to as *Stein factors* or *magic factors*.

In the case of (4.1) we can prove for the supremum norm that

where the last bound is of course only applicable if is differentiable (or at least Lipschitz-continuous, which, for example, is not the case if we regard the total variation metric or the Kolmogorov metric!). As the standard normal distribution has no extra parameters, in this specific case, the constants are free of additional parameters.

Note that, up to this point, we did not make use of the random variable . So, the steps up to here in general have to be calculated only once for a specific combination of distribution , metric and Stein operator . However, if we have bounds in the general form (5.1), we usually are able to treat many probability metrics together. Furthermore as there is often a particular 'good' Stein operator for a distribution (e.g., no other operator than (2.3) has been used for the standard normal distribution up to now), one can often just start with the next step below, if bounds of the form (5.1) are already available (which is the case for many distributions).

### An abstract approximation theorem

We are now in a position to bound the left hand side of (3.1). As this step heavily depends on the form of the Stein operator, we directly regard the case of the standard normal distribution.

Now, at this point we could directly plug in our random variable which we want to approximate and try to find upper bounds. However, it is often fruitful to formulate a more general theorem using only abstract properties of . Let us consider here the case of local dependence.

To this end, assume that is a sum of random variables such that the and variance . Assume that, for every , there is a set , such that is independent of all the random variables with . We call this set the 'neighborhood' of . Likewise let be a set such that all with are independent of all , . We can think of as the neighbors in the neighborhood of , a second-order neighborhood, so to speak. For a set define now the sum .

Using basically only Taylor expansion, it is possible to prove that

Note that, if we follow this line of argument, we can bound (1.1) only for functions where is bounded because of the third inequality of (5.2) (and in fact, if has discontinuities, so will ). To obtain a bound similar to (6.1) which contains only the expressions and , the argument is much more involved and the result is not as simple as (6.1); however, it can be done.

**Theorem A**. If is as described above, we have for the Lipschitz metric that

**Proof**. Recall that the Lipschitz metric is of the form (1.1) where the functions are Lipschitz-continuous with Lipschitz-constant 1, thus . Combining this with (6.1) and the last bound in (5.2) proves the theorem.

Thus, roughly speaking, we have proved that, to calculate the Lipschitz-distance between a with local dependence structure and a standard normal distribution, we only need to know the third moments of and the size of the neighborhoods and .

### Application of the theorem

We can treat the case of sums of independent and identically distributed random variables with Theorem A. So assume now that , and . We can take and we obtain from Theorem A that

## Connections to other methods

**Lindeberg's method**. Lindeberg (1922) introduced in a seminal article a method, where the difference in (1.1) is directly bounded. This method usually also heavily relies on Taylor expansion and thus shows some similarities with Stein's method.

**Tikhomirov's method**. Clearly the approach via (1.1) and (3.1) does not involve characteristic functions. However, Tikhomirov (1980) presented a proof of a central limit theorem based on characteristic functions and a differential operator similar to (2.3). The basic observation is that the characteristic function of the standard normal distribution satisfies the differential equation for all . Thus, if the characteristic function of is such that we expect that and hence that is close to the normal distribution. Tikhomirov states in his paper that he was inspired by Stein's seminal paper.

## Notes

- ↑
^{1.0}^{1.1}{{#invoke:Citation/CS1|citation |CitationClass=journal }} - ↑ Charles Stein: The Invariant, the Direct and the "Pretentious". Interview given in 2003 in Singapore
- ↑ {{#invoke:Citation/CS1|citation |CitationClass=journal }}

## References

- {{#invoke:Citation/CS1|citation

|CitationClass=journal }}

- {{#invoke:Citation/CS1|citation

|CitationClass=journal }}

- {{#invoke:Citation/CS1|citation

|CitationClass=journal }}

- {{#invoke:Citation/CS1|citation

|CitationClass=journal }}

- {{#invoke:Citation/CS1|citation

|CitationClass=journal }}

- {{#invoke:Citation/CS1|citation

|CitationClass=journal }}

- {{#invoke:Citation/CS1|citation

|CitationClass=journal }}

- {{#invoke:citation/CS1|citation

|CitationClass=book }}

- {{#invoke:citation/CS1|citation

|CitationClass=book }}

- {{#invoke:Citation/CS1|citation

|CitationClass=journal }}

## Literature

The following text is advanced, and gives a comprehensive overview of the normal case

- {{#invoke:citation/CS1|citation

|CitationClass=book }}

Another advanced book, but having some introductory character, is

- {{#invoke:citation/CS1|citation

|CitationClass=book }}

A standard reference is the book by Stein,

- {{#invoke:citation/CS1|citation

|CitationClass=book }} which contains a lot of interesting material, but may be a little hard to understand at first reading.

Despite its age, there are few standard introductory books about Stein's method available. The following recent textbook has a chapter (Chapter 2) devoted to introducing Stein's method:

- {{#invoke:citation/CS1|citation

|CitationClass=book }}

Although the book

- {{#invoke:citation/CS1|citation

|CitationClass=book }} is by large parts about Poisson approximation, it contains nevertheless a lot of information about the generator approach, in particular in the context of Poisson process approximation.

The following textbook has a chapter (Chapter 10) devoted to introducing Stein's method of Poisson approximation:

- {{#invoke:citation/CS1|citation

|CitationClass=book }}